feat(code): auto-assign PRs created by code agent to triggering human

[ifireball]

What happens

When the code agent opens a pull request, the PR author is the fullsend-ai-coder GitHub App. GitHub does not allow assigning that app user on PRs in our repos (replaceActorsForAssignable fails for app/fullsend-ai-coder and fullsend-ai-coder[bot]).

The PR therefore stays unassigned unless someone assigns it manually. In practice, many agent PRs sit in the backlog with no assignee, which makes ownership unclear for review and triage.

This is separate from #1838, which proposes assigning the issue to the coding agent. This issue is about assigning the PR to a human using the same ownership signals we use when triaging manually.

What should happen

After post-code.sh creates or updates a PR for issue #N, the post-script should assign the PR to a human GitHub user using this precedence (first match wins):

1. Person who invoked /fs-code on the issue (when the run was triggered by that slash command)
2. Issue assignee (first human assignee, if any)
3. Issue reporter (human issue author)

Rules:

• Never assign bots or GitHub Apps (app/*, *[bot], dependabot, etc.).
• If no human matches the precedence chain, leave the PR unassigned (do not guess from unrelated PRs or retro context).
• When updating an existing PR branch (PR already exists), apply the same assignment logic if the PR has no assignees yet.
• Assignment must be best-effort: log and continue if gh pr edit --add-assignee fails (e.g. user lacks permission), without failing the workflow.

Context

Manual triage on fullsend-ai/fullsend (2026-06-04) used this precedence on ~30 unassigned PRs. Human-authored PRs map cleanly to their authors; bot-authored PRs need the chain above. Examples that worked well:

• /fs-code dispatcher → PR assignee (e.g. issue reconcile-repos: enrollment commits lack a body, causing gitlint failures in onboarded repos #1824 → PR fix(#1824): add body to enrollment commit messages #1825)
• Issue author when human (e.g. issue /fs-review on a merged PR gives no feedback #1827 → PR fix(#1827): post feedback comment when /fs-review targets non-open PR #1832)
• Issue author for retro-filed fixes (e.g. issues Review agent: post inline comments for comment-only verdicts #1787, sandbox.Upload() silently drops files above ~16MB #1699 → PRs fix(#1787): post inline comments for COMMENT review verdicts #1792, fix(#1699): add post-upload verification to sandbox.Upload() #1702)

Cases that should remain unassigned under this policy: retro-only issues with no human author, assignee, or /fs-code invoker (e.g. issues filed by app/fullsend-ai-retro with no human comment).

Implementation notes (for implementers)

This should be implementable deterministically in post-code.sh (no LLM), immediately after gh pr create / when reusing an existing PR:

• post-code.sh already has ISSUE_NUMBER, REPO_FULL_NAME, PUSH_TOKEN, and gh access.
• Issue assignee and author are available via gh issue view "${ISSUE_NUMBER}" --json assignees,author.
• /fs-fix already plumbs TRIGGER_SOURCE from dispatch through reusable-fix.yml into its post-script; /fs-code does not set TRIGGER_SOURCE today even though dispatch has COMMENT_USER_LOGIN for issue_comment events. Consider mirroring the fix-agent plumbing (dispatch.yml → reusable-dispatch.yml → reusable-code.yml → harness/code.yaml runner_env) so label-triggered runs fall through to assignee/reporter only.
• For ready-to-code label triggers, precedence steps 2–3 apply; step 1 is N/A unless TRIGGER_SOURCE is also set for authorized label appliers in a follow-up.

Relevant files:

• internal/scaffold/fullsend-repo/scripts/post-code.sh (PR create path ~section 8)
• internal/scaffold/fullsend-repo/harness/code.yaml (runner_env)
• internal/scaffold/fullsend-repo/.github/workflows/dispatch.yml (/fs-code routing)
• .github/workflows/reusable-code.yml and reusable-dispatch.yml

Add shell tests in post-code-test.sh for the resolver (bot filtering, precedence, no-op when all candidates are bots).

[fullsend-ai-triage]

Triage Summary

Severity: medium · Category: feature

Problem

PRs opened by the code agent are authored by the fullsend-ai-coder App, which GitHub cannot assign. This leaves agent PRs without an assignee, making review ownership unclear.

Recommended Approach

Add a human-assignee resolver to post-code.sh (after PR create/reuse) with this precedence:

1. /fs-code invoker (via TRIGGER_SOURCE / COMMENT_USER_LOGIN)
2. First human issue assignee
3. Human issue author

Bot accounts (app/*, *[bot], dependabot) must be filtered out. Assignment should be best-effort — log and continue on failure.

The /fs-code dispatch path needs TRIGGER_SOURCE plumbed through dispatch.yml → reusable-dispatch.yml → reusable-code.yml → harness/code.yaml runner_env, mirroring the existing fix-agent pattern.

Key Files

• internal/scaffold/fullsend-repo/scripts/post-code.sh — PR create path
• internal/scaffold/fullsend-repo/harness/code.yaml — runner_env
• .github/workflows/dispatch.yml and reusable-code.yml — dispatch plumbing

Proposed Test Case

# post-code-test.sh additions
# 1. Human TRIGGER_SOURCE → selected as assignee
# 2. Bot TRIGGER_SOURCE + human issue assignee → assignee wins
# 3. All candidates are bots → no assignment attempted
# 4. PR already has assignees → no reassignment
# 5. gh pr edit failure → logged, script continues

This is well-specified and ready for implementation. Related #1838 (issue assignment) is a separate concern.

---

Labels: Adding component/dispatch since implementation requires plumbing TRIGGER_SOURCE through dispatch workflows, and priority/medium for non-urgent operational improvement.

[fullsend-ai-prioritize]

RICE Priority Score: 1.35

Score breakdown

Dimension	Score	Reasoning
Reach	1.5	Affects all users who interact with code-agent PRs. The issue notes ~30 unassigned PRs from a single triage session, indicating this is a recurring problem across multiple team members and customers who trigger /fs-code runs. Scored 1.5 for multiple affected users across the org.
Impact	1	The impact is a workflow inconvenience — PRs sit unassigned making triage and review ownership unclear. However, a simple manual workaround exists (assign the PR by hand). This is a noticeable improvement to the triage workflow but not blocking. Scored 1 for medium impact.
Confidence	0.9	Very well-specified issue with clear precedence rules, specific files listed, existing patterns to mirror (fix-agent plumbing), and concrete test cases. The scope is well-bounded. Scored 0.9.
Effort	1	Deterministic shell changes in post-code.sh with clear implementation path mirroring existing fix-agent patterns. Touches 4-5 files (post-code.sh, code.yaml, dispatch workflows) plus test additions. Well-scoped but requires plumbing TRIGGER_SOURCE through multiple workflow layers. Scored 1 for medium effort.

Formula: (1.5 x 1 x 0.9) / 1 = 1.35

[ifireball]

/fs-code